ADHTransNet basado en radiómica en RM multimodal de hipófisis para predicción hormonal no invasiva en niños

Growth hormone deficiency (GHD) and idiopathic central precocious puberty (ICPP) are typically diagnosed through invasive stimulation tests that require multiple blood samples collected over time. To reduce the need for such procedures, the study aims to establish an adjunctive tool by devising a fully automated pipeline for adenohypophysis segmentation and radiomics-based prediction of growth hormone (arg-pGH and ins-pGH in GHD) and gonadotropin (pLH and pLH/FSH in ICPP) levels in children.

A total of 274 subjects with 548 scans (T1-weighted and T2-weighted images, T1WI and T2WI) were identified, including GHD, ICPP, and normal control groups. MRI acquisition was performed 1 day prior to the hormone stimulation tests. The automated segmentation of adenohypophysis (ADH) on pituitary MRI was first achieved by the proposed ADHTransNet. Then, the radiomics features were extracted, and the consistency was assessed between manual and automated segmentations. Lastly, using a full-search feature selection strategy, we developed radiomics-based models to predict arginine-stimulated growth hormone (arg-pGH) and insulin-stimulated growth hormone (ins-pGH) levels in patients with GHD, as well as luteinizing hormone (pLH) levels and the pLH/FSH ratio in patients with ICPP.

The superior ADH segmentation was achieved by ADHTransNet over other deep learning methods under comparison. The radiomics was validated with high measurement consistency and statistical consistency of the statistical T-values on both T1WI and T2WI images. Significant correlations were observed between truth hormone level and the predicted the peak GH of arginine stimulation test in GHD group (r=0.422, p<0.001), the peak GH of insulin stimulation test in GHD group (r=0.359, p<0.001), the peak luteinizing hormone (LH) in ICPP group (r=0.680, p<0.001), and the ratio of peak LH to peak follicle-stimulating hormone (FSH) in ICPP group(r=0.766, p<0.001).

This fully automated, multimodal, reproducible, and non-invasive pipeline shows promise in predicting GH and gonadotropin levels from MRI, reducing reliance on repeated blood tests, and enhancing assessment of hormone-related disorders.